Loading...
Apple introduces SimpleDesign, an AI model for protein design
SimpleDesign can generate protein sequences and structures

Apple introduces SimpleDesign, an AI model for protein design

Sep 13, 2026
11:17 am

What's the story

Apple has unveiled a new artificial intelligence (AI) model named SimpleDesign. The innovative system can simultaneously generate both the sequences and structures of proteins. The announcement comes as a continuation of Apple's research in the field, following their previous study on SimpleFold, a model that predicts a protein's 3D structure from its amino acid sequence using flow-matching techniques.

Model innovation

How is SimpleDesign different from existing models?

Unlike existing models that depend on a multi-stage training process, SimpleDesign learns to generate amino acid sequences and continuous 3D structures in one go.

It does this by directly learning from paired amino acid sequences and 3D coordinates, instead of first compressing protein structures into a separate tokenized representation.

This makes SimpleDesign an effective multi-modal protein design model trained directly in the data space.

Training details

Training process of SimpleDesign

The team trained SimpleDesign with over two million protein sequence-and-structure pairs, mostly from the AFESM dataset.

They corrupted both parts of each pair during training: amino acids in the sequence were randomly hidden behind masked tokens while noise was added to the corresponding 3D structure.

This way, they could simulate different scenarios like protein folding or inverse folding and effectively train SimpleDesign for protein co-design tasks.

ADVERTISEMENT

Model evaluation

Performance of the new model

Despite a simpler training pipeline, SimpleDesign performed competitively on protein co-design, structure generation, and sequence generation benchmarks.

The model could generate plausible protein structures with amino acid sequences that were as good or better than those from most competing multimodal models.

However, it's worth noting that these results are still limited to computer-based evaluations as the generated proteins weren't experimentally tested for real-world biological systems.

ADVERTISEMENT